Introduction to the Tiled HW Architecture of SHAPES

Nanoscale systems on chip dedicated to embedded systems and numerical computations will integrate a few hundreds of million gates. The challenge is to find a scalable HW/SW design style for future CMOS technologies. The main HW problem is wiring, which threatens Moore’s law. Tiled architectures suggest a possible HW path: “small” processing tiles connected by “short wires”. A second HW problem is the management of the design complexity. A tiled design style reuses stable Intellectual Properties requiring a few million gates: a manageable complexity. A typical SHAPES tile always contains one Distributed Network Processor (DNP) for inter-tile communications, plus one VLIW DSP processor for computation and/or one RISC processor for control. The DNP is supported by a NoC (for inter-tile, inter-chip communications) and by N-dimensional toroidal network for off-chip communications. The SW challenge is to provide a simple and efficient programming environment for a (massive) tiled parallel architecture. This paper introduces the HW architecture.